Use while loop instead of for-range loop in broadcast kernel (#448)

This reduces the overhead incurred by the broadcast kernel over a
handwritten kernel by 10-15% on M1 with Metal.jl.

Author: maxwindiff

src/host/broadcast.jl

@@ -52,7 +52,9 @@ end
 
     # grid-stride kernel
     function broadcast_kernel(ctx, dest, bc′, nelem)
-        for i in 1:nelem
+        i = 0
+        while i < nelem
+            i += 1
             I = @cartesianidx(dest, i)
             @inbounds dest[I] = bc′[I]
         end